A badge under the chart reads: "price rose in 16 of 28 similar cases." The most damaging way to read it is as "57 percent probability of a rise."
This page is about what sits behind that number, and why the word "probability" does not appear anywhere near it on our pages.
What was actually counted
Twenty-eight places in the instrument's history matched the current move at the chosen threshold. For each, the stretch after the match was taken and the endpoint compared to the start. In sixteen cases it finished higher.
That is the whole claim. It is entirely about the past and contains no assumption about the future. It also describes a sample you defined yourself with the sliders: change the similarity threshold and both the match count and the outcome split change with it.
The horizon depends on the timeframe
"What happened next" — over how long?
| Timeframe | Horizon | In wall-clock terms |
|---|---|---|
| 1 hour | 100 candles | about four days |
| 4 hours | 50 candles | about eight days |
| 1 day | 30 candles | about six weeks |
The horizons differ on purpose. Thirty candles on an hourly chart is a day and a half — too short for an outcome to mean anything. A hundred candles on a daily chart is nearly half a year, over which everything changes.
The consequence is easy to miss: the same instrument on different timeframes is answering different questions. Outcome shares are not comparable across them.
The regime filter
Each match has a context: was the market broadly above or below its long moving average. Matches whose context contradicts the present one are removed.
Technically this is price relative to a 200-period moving average. If fewer than two hundred candles of history existed at that point, the nearest available period is used — 100, 50 or 20.
The filter is deliberately crude and does exactly one thing: it stops an episode from a long uptrend being compared with one from a long downtrend. It does not make 2018 comparable to 2026. Market structure, participants, liquidity, regulation — it sees none of that.
Four limits worth knowing
The sample is small. Twenty-eight episodes is twenty-eight. At that size the difference between "16 of 28" and "12 of 28" is comfortably explained by chance. Quoting such a share to a third decimal is theatre.
Episodes are not independent. Overlapping windows are removed, so no single stretch of history appears twice. But matches still cluster in time: if the market spent a week in March 2021 moving this way, several nearby episodes come from there and carry roughly the same information.
History kept the survivors. The instrument list contains what is still trading and has a long price record. Coins that went to zero and companies that were delisted never enter the search, because their data is not there. Any statistic over such a sample leans toward benign outcomes.
Multiple testing. The similarity threshold is a dial controlling how many matches exist. On BTC/USDT 4H, measured 25 September 2026, threshold 90 returned zero matches, 80 returned twenty-eight, 70 returned fifty-three. Turn that dial until the picture becomes convincing and what you have made convincing is the picture.
Why we do not print a probability
A probability is a claim about the process generating the data. Making one requires a model of that process and grounds to believe past observations were drawn from it.
We have neither. We have stretches of chart with a similar outline and a record of what followed them. Converting a share into a probability would quietly smuggle in the assumption that markets repeat, and we do not make that assumption.
For the same reason our pages carry no phrasing like "price will return" or "the zone will play out". The markup describes what already happened and what follows from it mechanically. Nothing beyond that.
How to use it honestly
Analogue statistics are useful for forming a question, not for receiving an answer.
Look at the matches themselves, not only the summary. Twenty-eight episodes is twenty-eight charts, and studying three or four of them tells you more than the outcome share. Often they turn out to have resolved in wildly different ways — which is itself the answer: this situation is not informative.
Watch the spread, not just the direction. If sixteen of twenty-eight ended higher but the rises were half a percent while the falls were ten, the outcome share is actively misleading.
And keep the threshold in mind. Output at 70 and output at 90 are not "more data and less data" — they are samples of different quality.
In short
The "N of M" badge describes the past over a sample you selected. The horizon depends on the timeframe, the context filter is coarse, the sample is small and survivor-biased, and you control its size yourself.
It is an honest instrument for looking at history. It does not become a forecast at any threshold.